Exploring Sentence Type Effects on the Lombard Effect and Intelligibility Enhancement: A Comparative Study of Natural and Grid Sentences

TL;DR

This paper investigates how sentence type (natural vs. grid) influences the Lombard effect and intelligibility enhancement. By constructing two closely matched corpora (LCT for natural Mandarin and EMALG for grid Mandarin) and applying StarGAN-based normal-to-Lombard conversion trained on each corpus, the authors quantify differences in phonetic/acoustic changes and intelligibility outcomes. The study finds grid sentences exhibit a more pronounced Lombard effect and yield higher objective intelligibility, while natural sentences better preserve speech quality after conversion. These findings inform the design of speech intelligibility enhancement systems for noisy environments, highlighting a trade-off between intelligibility gains and perceived naturalness depending on sentence type.

Abstract

This study explores how sentence types affect the Lombard effect and intelligibility enhancement, focusing on comparisons between natural and grid sentences. Using the Lombard Chinese-TIMIT (LCT) corpus and the Enhanced MAndarin Lombard Grid (EMALG) corpus, we analyze changes in phonetic and acoustic features across different noise levels. Our results show that grid sentences produce more pronounced Lombard effects than natural sentences. Then, we develop and test a normal-to-Lombard conversion model, trained separately on LCT and EMALG corpora. Through subjective and objective evaluations, natural sentences are superior in maintaining speech quality in intelligibility enhancement. In contrast, grid sentences could provide superior intelligibility due to the more pronounced Lombard effect. This study provides a valuable perspective on enhancing speech communication in noisy environments.

Figures (5)

• Figure 1: Block diagram of the comparative study.
• Figure 2: Our implemented StarGAN architecture with six domains for intelligibility enhancement. In inference, normal speech are converted to L2 speech.
• Figure 3: Lombard effect analysis 1. The mean and standard deviation of vowel duration, vowel ratio, F1, and F2 of male (M) speakers, female (F) speakers and all (ALL) speakers. Error bars represent standard deviation. * means a significant difference (p $<$ 0.001) in t-test between two different speaking styles across genders.
• Figure 4: Lombard effect analysis 2. The mean and standard deviation of F0, loudness, and alpha ratio.
• Figure 5: Boxplots of subjective evaluation results for LCT-StarGAN and EMALG-StarGAN. Higher is better. T1-T12 correspond to 12 pairs of test sequence.